perf: disable kvcache for semantic by default #368
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
Author
|
Hm, nevermind here, I think I drew an invalid conclusion due to a mixup. It seems that caching still results in a small performance gain on the text model. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
An analysis of my previous patches suggests that use_kv_caching=True is not a win at all for the text model.
The current KV caching code has quite a bit of overhead. In particular, it (re)allocates large tensors with adding to the cache via a
torch.catcall. With the coarse model it seems to result in some performance gain, but it is definitely not a win for the smaller text model. It seems that the KV cache was what was causing the bimodality in #366, I'm guessing sometimes we would get unlucky and the model would have to reallocate.With this change I can consistently get about 280 it/second performance for the text model on an H100 after warmup.